Chameleon - Image Style Transfer Based on Image Classification Networks.

In recent years, deep neural networks enabled computers to extract features of images; this stimulated interest in image style transfer. Since the first NST algorithm, many schemes are proposed to accelerate the process. Yet, these methods more or less sacrifice the visual quality of the output. In this paper, we analyze two precursive algorithms and generalize the results. We propose an Image Style Transfer method based on Image Classification Networks. With Chameleon, we accelerate the runtime by 11 times and we show that general CNN techniques would improve the stylization process. Our work provides novel insights into the classification capacity of CNNs and demonstrates the potential of NST algorithms.